drivers/block/null_blk.c

   1 #include <linux/module.h>
   2
   3 #include <linux/moduleparam.h>
   4 #include <linux/sched.h>
   5 #include <linux/fs.h>
   6 #include <linux/blkdev.h>
   7 #include <linux/init.h>
   8 #include <linux/slab.h>
   9 #include <linux/blk-mq.h>
  10 #include <linux/hrtimer.h>
  11
  12 struct nullb_cmd {
  13         struct list_head list;
  14         struct llist_node ll_list;
  15         struct call_single_data csd;
  16         struct request *rq;
  17         struct bio *bio;
  18         unsigned int tag;
  19         struct nullb_queue *nq;
  20 };
  21
  22 struct nullb_queue {
  23         unsigned long *tag_map;
  24         wait_queue_head_t wait;
  25         unsigned int queue_depth;
  26
  27         struct nullb_cmd *cmds;
  28 };
  29
  30 struct nullb {
  31         struct list_head list;
  32         unsigned int index;
  33         struct request_queue *q;
  34         struct gendisk *disk;
  35         struct blk_mq_tag_set tag_set;
  36         struct hrtimer timer;
  37         unsigned int queue_depth;
  38         spinlock_t lock;
  39
  40         struct nullb_queue *queues;
  41         unsigned int nr_queues;
  42 };
  43
  44 static LIST_HEAD(nullb_list);
  45 static struct mutex lock;
  46 static int null_major;
  47 static int nullb_indexes;
  48
  49 struct completion_queue {
  50         struct llist_head list;
  51         struct hrtimer timer;
  52 };
  53
  54 /*
  55  * These are per-cpu for now, they will need to be configured by the
  56  * complete_queues parameter and appropriately mapped.
  57  */
  58 static DEFINE_PER_CPU(struct completion_queue, completion_queues);
  59
  60 enum {
  61         NULL_IRQ_NONE           = 0,
  62         NULL_IRQ_SOFTIRQ        = 1,
  63         NULL_IRQ_TIMER          = 2,
  64 };
  65
  66 enum {
  67         NULL_Q_BIO              = 0,
  68         NULL_Q_RQ               = 1,
  69         NULL_Q_MQ               = 2,
  70 };
  71
  72 static int submit_queues;
  73 module_param(submit_queues, int, S_IRUGO);
  74 MODULE_PARM_DESC(submit_queues, "Number of submission queues");
  75
  76 static int home_node = NUMA_NO_NODE;
  77 module_param(home_node, int, S_IRUGO);
  78 MODULE_PARM_DESC(home_node, "Home node for the device");
  79
  80 static int queue_mode = NULL_Q_MQ;
  81
  82 static int null_param_store_val(const char *str, int *val, int min, int max)
  83 {
  84         int ret, new_val;
  85
  86         ret = kstrtoint(str, 10, &new_val);
  87         if (ret)
  88                 return -EINVAL;
  89
  90         if (new_val < min || new_val > max)
  91                 return -EINVAL;
  92
  93         *val = new_val;
  94         return 0;
  95 }
  96
  97 static int null_set_queue_mode(const char *str, const struct kernel_param *kp)
  98 {
  99         return null_param_store_val(str, &queue_mode, NULL_Q_BIO, NULL_Q_MQ);
 100 }
 101
 102 static const struct kernel_param_ops null_queue_mode_param_ops = {
 103         .set    = null_set_queue_mode,
 104         .get    = param_get_int,
 105 };
 106
 107 device_param_cb(queue_mode, &null_queue_mode_param_ops, &queue_mode, S_IRUGO);
 108 MODULE_PARM_DESC(queue_mode, "Block interface to use (0=bio,1=rq,2=multiqueue)");
 109
 110 static int gb = 250;
 111 module_param(gb, int, S_IRUGO);
 112 MODULE_PARM_DESC(gb, "Size in GB");
 113
 114 static int bs = 512;
 115 module_param(bs, int, S_IRUGO);
 116 MODULE_PARM_DESC(bs, "Block size (in bytes)");
 117
 118 static int nr_devices = 2;
 119 module_param(nr_devices, int, S_IRUGO);
 120 MODULE_PARM_DESC(nr_devices, "Number of devices to register");
 121
 122 static int irqmode = NULL_IRQ_SOFTIRQ;
 123
 124 static int null_set_irqmode(const char *str, const struct kernel_param *kp)
 125 {
 126         return null_param_store_val(str, &irqmode, NULL_IRQ_NONE,
 127                                         NULL_IRQ_TIMER);
 128 }
 129
 130 static const struct kernel_param_ops null_irqmode_param_ops = {
 131         .set    = null_set_irqmode,
 132         .get    = param_get_int,
 133 };
 134
 135 device_param_cb(irqmode, &null_irqmode_param_ops, &irqmode, S_IRUGO);
 136 MODULE_PARM_DESC(irqmode, "IRQ completion handler. 0-none, 1-softirq, 2-timer");
 137
 138 static int completion_nsec = 10000;
 139 module_param(completion_nsec, int, S_IRUGO);
 140 MODULE_PARM_DESC(completion_nsec, "Time in ns to complete a request in hardware. Default: 10,000ns");
 141
 142 static int hw_queue_depth = 64;
 143 module_param(hw_queue_depth, int, S_IRUGO);
 144 MODULE_PARM_DESC(hw_queue_depth, "Queue depth for each hardware queue. Default: 64");
 145
 146 static bool use_per_node_hctx = false;
 147 module_param(use_per_node_hctx, bool, S_IRUGO);
 148 MODULE_PARM_DESC(use_per_node_hctx, "Use per-node allocation for hardware context queues. Default: false");
 149
 150 static void put_tag(struct nullb_queue *nq, unsigned int tag)
 151 {
 152         clear_bit_unlock(tag, nq->tag_map);
 153
 154         if (waitqueue_active(&nq->wait))
 155                 wake_up(&nq->wait);
 156 }
 157
 158 static unsigned int get_tag(struct nullb_queue *nq)
 159 {
 160         unsigned int tag;
 161
 162         do {
 163                 tag = find_first_zero_bit(nq->tag_map, nq->queue_depth);
 164                 if (tag >= nq->queue_depth)
 165                         return -1U;
 166         } while (test_and_set_bit_lock(tag, nq->tag_map));
 167
 168         return tag;
 169 }
 170
 171 static void free_cmd(struct nullb_cmd *cmd)
 172 {
 173         put_tag(cmd->nq, cmd->tag);
 174 }
 175
 176 static struct nullb_cmd *__alloc_cmd(struct nullb_queue *nq)
 177 {
 178         struct nullb_cmd *cmd;
 179         unsigned int tag;
 180
 181         tag = get_tag(nq);
 182         if (tag != -1U) {
 183                 cmd = &nq->cmds[tag];
 184                 cmd->tag = tag;
 185                 cmd->nq = nq;
 186                 return cmd;
 187         }
 188
 189         return NULL;
 190 }
 191
 192 static struct nullb_cmd *alloc_cmd(struct nullb_queue *nq, int can_wait)
 193 {
 194         struct nullb_cmd *cmd;
 195         DEFINE_WAIT(wait);
 196
 197         cmd = __alloc_cmd(nq);
 198         if (cmd || !can_wait)
 199                 return cmd;
 200
 201         do {
 202                 prepare_to_wait(&nq->wait, &wait, TASK_UNINTERRUPTIBLE);
 203                 cmd = __alloc_cmd(nq);
 204                 if (cmd)
 205                         break;
 206
 207                 io_schedule();
 208         } while (1);
 209
 210         finish_wait(&nq->wait, &wait);
 211         return cmd;
 212 }
 213
 214 static void end_cmd(struct nullb_cmd *cmd)
 215 {
 216         switch (queue_mode)  {
 217         case NULL_Q_MQ:
 218                 blk_mq_end_request(cmd->rq, 0);
 219                 return;
 220         case NULL_Q_RQ:
 221                 INIT_LIST_HEAD(&cmd->rq->queuelist);
 222                 blk_end_request_all(cmd->rq, 0);
 223                 break;
 224         case NULL_Q_BIO:
 225                 bio_endio(cmd->bio, 0);
 226                 break;
 227         }
 228
 229         free_cmd(cmd);
 230 }
 231
 232 static enum hrtimer_restart null_cmd_timer_expired(struct hrtimer *timer)
 233 {
 234         struct completion_queue *cq;
 235         struct llist_node *entry;
 236         struct nullb_cmd *cmd;
 237
 238         cq = &per_cpu(completion_queues, smp_processor_id());
 239
 240         while ((entry = llist_del_all(&cq->list)) != NULL) {
 241                 entry = llist_reverse_order(entry);
 242                 do {
 243                         struct request_queue *q = NULL;
 244
 245                         cmd = container_of(entry, struct nullb_cmd, ll_list);
 246                         entry = entry->next;
 247                         if (cmd->rq)
 248                                 q = cmd->rq->q;
 249                         end_cmd(cmd);
 250
 251                         if (q && !q->mq_ops && blk_queue_stopped(q)) {
 252                                 spin_lock(q->queue_lock);
 253                                 if (blk_queue_stopped(q))
 254                                         blk_start_queue(q);
 255                                 spin_unlock(q->queue_lock);
 256                         }
 257                 } while (entry);
 258         }
 259
 260         return HRTIMER_NORESTART;
 261 }
 262
 263 static void null_cmd_end_timer(struct nullb_cmd *cmd)
 264 {
 265         struct completion_queue *cq = &per_cpu(completion_queues, get_cpu());
 266
 267         cmd->ll_list.next = NULL;
 268         if (llist_add(&cmd->ll_list, &cq->list)) {
 269                 ktime_t kt = ktime_set(0, completion_nsec);
 270
 271                 hrtimer_start(&cq->timer, kt, HRTIMER_MODE_REL_PINNED);
 272         }
 273
 274         put_cpu();
 275 }
 276
 277 static void null_softirq_done_fn(struct request *rq)
 278 {
 279         if (queue_mode == NULL_Q_MQ)
 280                 end_cmd(blk_mq_rq_to_pdu(rq));
 281         else
 282                 end_cmd(rq->special);
 283 }
 284
 285 static inline void null_handle_cmd(struct nullb_cmd *cmd)
 286 {
 287         /* Complete IO by inline, softirq or timer */
 288         switch (irqmode) {
 289         case NULL_IRQ_SOFTIRQ:
 290                 switch (queue_mode)  {
 291                 case NULL_Q_MQ:
 292                         blk_mq_complete_request(cmd->rq);
 293                         break;
 294                 case NULL_Q_RQ:
 295                         blk_complete_request(cmd->rq);
 296                         break;
 297                 case NULL_Q_BIO:
 298                         /*
 299                          * XXX: no proper submitting cpu information available.
 300                          */
 301                         end_cmd(cmd);
 302                         break;
 303                 }
 304                 break;
 305         case NULL_IRQ_NONE:
 306                 end_cmd(cmd);
 307                 break;
 308         case NULL_IRQ_TIMER:
 309                 null_cmd_end_timer(cmd);
 310                 break;
 311         }
 312 }
 313
 314 static struct nullb_queue *nullb_to_queue(struct nullb *nullb)
 315 {
 316         int index = 0;
 317
 318         if (nullb->nr_queues != 1)
 319                 index = raw_smp_processor_id() / ((nr_cpu_ids + nullb->nr_queues - 1) / nullb->nr_queues);
 320
 321         return &nullb->queues[index];
 322 }
 323
 324 static void null_queue_bio(struct request_queue *q, struct bio *bio)
 325 {
 326         struct nullb *nullb = q->queuedata;
 327         struct nullb_queue *nq = nullb_to_queue(nullb);
 328         struct nullb_cmd *cmd;
 329
 330         cmd = alloc_cmd(nq, 1);
 331         cmd->bio = bio;
 332
 333         null_handle_cmd(cmd);
 334 }
 335
 336 static int null_rq_prep_fn(struct request_queue *q, struct request *req)
 337 {
 338         struct nullb *nullb = q->queuedata;
 339         struct nullb_queue *nq = nullb_to_queue(nullb);
 340         struct nullb_cmd *cmd;
 341
 342         cmd = alloc_cmd(nq, 0);
 343         if (cmd) {
 344                 cmd->rq = req;
 345                 req->special = cmd;
 346                 return BLKPREP_OK;
 347         }
 348         blk_stop_queue(q);
 349
 350         return BLKPREP_DEFER;
 351 }
 352
 353 static void null_request_fn(struct request_queue *q)
 354 {
 355         struct request *rq;
 356
 357         while ((rq = blk_fetch_request(q)) != NULL) {
 358                 struct nullb_cmd *cmd = rq->special;
 359
 360                 spin_unlock_irq(q->queue_lock);
 361                 null_handle_cmd(cmd);
 362                 spin_lock_irq(q->queue_lock);
 363         }
 364 }
 365
 366 static int null_queue_rq(struct blk_mq_hw_ctx *hctx,
 367                          const struct blk_mq_queue_data *bd)
 368 {
 369         struct nullb_cmd *cmd = blk_mq_rq_to_pdu(bd->rq);
 370
 371         cmd->rq = bd->rq;
 372         cmd->nq = hctx->driver_data;
 373
 374         blk_mq_start_request(bd->rq);
 375
 376         null_handle_cmd(cmd);
 377         return BLK_MQ_RQ_QUEUE_OK;
 378 }
 379
 380 static void null_init_queue(struct nullb *nullb, struct nullb_queue *nq)
 381 {
 382         BUG_ON(!nullb);
 383         BUG_ON(!nq);
 384
 385         init_waitqueue_head(&nq->wait);
 386         nq->queue_depth = nullb->queue_depth;
 387 }
 388
 389 static int null_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 390                           unsigned int index)
 391 {
 392         struct nullb *nullb = data;
 393         struct nullb_queue *nq = &nullb->queues[index];
 394
 395         hctx->driver_data = nq;
 396         null_init_queue(nullb, nq);
 397         nullb->nr_queues++;
 398
 399         return 0;
 400 }
 401
 402 static struct blk_mq_ops null_mq_ops = {
 403         .queue_rq       = null_queue_rq,
 404         .map_queue      = blk_mq_map_queue,
 405         .init_hctx      = null_init_hctx,
 406         .complete       = null_softirq_done_fn,
 407 };
 408
 409 static void null_del_dev(struct nullb *nullb)
 410 {
 411         list_del_init(&nullb->list);
 412
 413         del_gendisk(nullb->disk);
 414         blk_cleanup_queue(nullb->q);
 415         if (queue_mode == NULL_Q_MQ)
 416                 blk_mq_free_tag_set(&nullb->tag_set);
 417         put_disk(nullb->disk);
 418         kfree(nullb);
 419 }
 420
 421 static int null_open(struct block_device *bdev, fmode_t mode)
 422 {
 423         return 0;
 424 }
 425
 426 static void null_release(struct gendisk *disk, fmode_t mode)
 427 {
 428 }
 429
 430 static const struct block_device_operations null_fops = {
 431         .owner =        THIS_MODULE,
 432         .open =         null_open,
 433         .release =      null_release,
 434 };
 435
 436 static int setup_commands(struct nullb_queue *nq)
 437 {
 438         struct nullb_cmd *cmd;
 439         int i, tag_size;
 440
 441         nq->cmds = kzalloc(nq->queue_depth * sizeof(*cmd), GFP_KERNEL);
 442         if (!nq->cmds)
 443                 return -ENOMEM;
 444
 445         tag_size = ALIGN(nq->queue_depth, BITS_PER_LONG) / BITS_PER_LONG;
 446         nq->tag_map = kzalloc(tag_size * sizeof(unsigned long), GFP_KERNEL);
 447         if (!nq->tag_map) {
 448                 kfree(nq->cmds);
 449                 return -ENOMEM;
 450         }
 451
 452         for (i = 0; i < nq->queue_depth; i++) {
 453                 cmd = &nq->cmds[i];
 454                 INIT_LIST_HEAD(&cmd->list);
 455                 cmd->ll_list.next = NULL;
 456                 cmd->tag = -1U;
 457         }
 458
 459         return 0;
 460 }
 461
 462 static void cleanup_queue(struct nullb_queue *nq)
 463 {
 464         kfree(nq->tag_map);
 465         kfree(nq->cmds);
 466 }
 467
 468 static void cleanup_queues(struct nullb *nullb)
 469 {
 470         int i;
 471
 472         for (i = 0; i < nullb->nr_queues; i++)
 473                 cleanup_queue(&nullb->queues[i]);
 474
 475         kfree(nullb->queues);
 476 }
 477
 478 static int setup_queues(struct nullb *nullb)
 479 {
 480         nullb->queues = kzalloc(submit_queues * sizeof(struct nullb_queue),
 481                                                                 GFP_KERNEL);
 482         if (!nullb->queues)
 483                 return -ENOMEM;
 484
 485         nullb->nr_queues = 0;
 486         nullb->queue_depth = hw_queue_depth;
 487
 488         return 0;
 489 }
 490
 491 static int init_driver_queues(struct nullb *nullb)
 492 {
 493         struct nullb_queue *nq;
 494         int i, ret = 0;
 495
 496         for (i = 0; i < submit_queues; i++) {
 497                 nq = &nullb->queues[i];
 498
 499                 null_init_queue(nullb, nq);
 500
 501                 ret = setup_commands(nq);
 502                 if (ret)
 503                         return ret;
 504                 nullb->nr_queues++;
 505         }
 506         return 0;
 507 }
 508
 509 static int null_add_dev(void)
 510 {
 511         struct gendisk *disk;
 512         struct nullb *nullb;
 513         sector_t size;
 514         int rv;
 515
 516         nullb = kzalloc_node(sizeof(*nullb), GFP_KERNEL, home_node);
 517         if (!nullb) {
 518                 rv = -ENOMEM;
 519                 goto out;
 520         }
 521
 522         spin_lock_init(&nullb->lock);
 523
 524         if (queue_mode == NULL_Q_MQ && use_per_node_hctx)
 525                 submit_queues = nr_online_nodes;
 526
 527         rv = setup_queues(nullb);
 528         if (rv)
 529                 goto out_free_nullb;
 530
 531         if (queue_mode == NULL_Q_MQ) {
 532                 nullb->tag_set.ops = &null_mq_ops;
 533                 nullb->tag_set.nr_hw_queues = submit_queues;
 534                 nullb->tag_set.queue_depth = hw_queue_depth;
 535                 nullb->tag_set.numa_node = home_node;
 536                 nullb->tag_set.cmd_size = sizeof(struct nullb_cmd);
 537                 nullb->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
 538                 nullb->tag_set.driver_data = nullb;
 539
 540                 rv = blk_mq_alloc_tag_set(&nullb->tag_set);
 541                 if (rv)
 542                         goto out_cleanup_queues;
 543
 544                 nullb->q = blk_mq_init_queue(&nullb->tag_set);
 545                 if (IS_ERR(nullb->q)) {
 546                         rv = -ENOMEM;
 547                         goto out_cleanup_tags;
 548                 }
 549         } else if (queue_mode == NULL_Q_BIO) {
 550                 nullb->q = blk_alloc_queue_node(GFP_KERNEL, home_node);
 551                 if (!nullb->q) {
 552                         rv = -ENOMEM;
 553                         goto out_cleanup_queues;
 554                 }
 555                 blk_queue_make_request(nullb->q, null_queue_bio);
 556                 rv = init_driver_queues(nullb);
 557                 if (rv)
 558                         goto out_cleanup_blk_queue;
 559         } else {
 560                 nullb->q = blk_init_queue_node(null_request_fn, &nullb->lock, home_node);
 561                 if (!nullb->q) {
 562                         rv = -ENOMEM;
 563                         goto out_cleanup_queues;
 564                 }
 565                 blk_queue_prep_rq(nullb->q, null_rq_prep_fn);
 566                 blk_queue_softirq_done(nullb->q, null_softirq_done_fn);
 567                 rv = init_driver_queues(nullb);
 568                 if (rv)
 569                         goto out_cleanup_blk_queue;
 570         }
 571
 572         nullb->q->queuedata = nullb;
 573         queue_flag_set_unlocked(QUEUE_FLAG_NONROT, nullb->q);
 574         queue_flag_clear_unlocked(QUEUE_FLAG_ADD_RANDOM, nullb->q);
 575
 576         disk = nullb->disk = alloc_disk_node(1, home_node);
 577         if (!disk) {
 578                 rv = -ENOMEM;
 579                 goto out_cleanup_blk_queue;
 580         }
 581
 582         mutex_lock(&lock);
 583         list_add_tail(&nullb->list, &nullb_list);
 584         nullb->index = nullb_indexes++;
 585         mutex_unlock(&lock);
 586
 587         blk_queue_logical_block_size(nullb->q, bs);
 588         blk_queue_physical_block_size(nullb->q, bs);
 589
 590         size = gb * 1024 * 1024 * 1024ULL;
 591         sector_div(size, bs);
 592         set_capacity(disk, size);
 593
 594         disk->flags |= GENHD_FL_EXT_DEVT | GENHD_FL_SUPPRESS_PARTITION_INFO;
 595         disk->major             = null_major;
 596         disk->first_minor       = nullb->index;
 597         disk->fops              = &null_fops;
 598         disk->private_data      = nullb;
 599         disk->queue             = nullb->q;
 600         sprintf(disk->disk_name, "nullb%d", nullb->index);
 601         add_disk(disk);
 602         return 0;
 603
 604 out_cleanup_blk_queue:
 605         blk_cleanup_queue(nullb->q);
 606 out_cleanup_tags:
 607         if (queue_mode == NULL_Q_MQ)
 608                 blk_mq_free_tag_set(&nullb->tag_set);
 609 out_cleanup_queues:
 610         cleanup_queues(nullb);
 611 out_free_nullb:
 612         kfree(nullb);
 613 out:
 614         return rv;
 615 }
 616
 617 static int __init null_init(void)
 618 {
 619         unsigned int i;
 620
 621         if (bs > PAGE_SIZE) {
 622                 pr_warn("null_blk: invalid block size\n");
 623                 pr_warn("null_blk: defaults block size to %lu\n", PAGE_SIZE);
 624                 bs = PAGE_SIZE;
 625         }
 626
 627         if (queue_mode == NULL_Q_MQ && use_per_node_hctx) {
 628                 if (submit_queues < nr_online_nodes) {
 629                         pr_warn("null_blk: submit_queues param is set to %u.",
 630                                                         nr_online_nodes);
 631                         submit_queues = nr_online_nodes;
 632                 }
 633         } else if (submit_queues > nr_cpu_ids)
 634                 submit_queues = nr_cpu_ids;
 635         else if (!submit_queues)
 636                 submit_queues = 1;
 637
 638         mutex_init(&lock);
 639
 640         /* Initialize a separate list for each CPU for issuing softirqs */
 641         for_each_possible_cpu(i) {
 642                 struct completion_queue *cq = &per_cpu(completion_queues, i);
 643
 644                 init_llist_head(&cq->list);
 645
 646                 if (irqmode != NULL_IRQ_TIMER)
 647                         continue;
 648
 649                 hrtimer_init(&cq->timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
 650                 cq->timer.function = null_cmd_timer_expired;
 651         }
 652
 653         null_major = register_blkdev(0, "nullb");
 654         if (null_major < 0)
 655                 return null_major;
 656
 657         for (i = 0; i < nr_devices; i++) {
 658                 if (null_add_dev()) {
 659                         unregister_blkdev(null_major, "nullb");
 660                         return -EINVAL;
 661                 }
 662         }
 663
 664         pr_info("null: module loaded\n");
 665         return 0;
 666 }
 667
 668 static void __exit null_exit(void)
 669 {
 670         struct nullb *nullb;
 671
 672         unregister_blkdev(null_major, "nullb");
 673
 674         mutex_lock(&lock);
 675         while (!list_empty(&nullb_list)) {
 676                 nullb = list_entry(nullb_list.next, struct nullb, list);
 677                 null_del_dev(nullb);
 678         }
 679         mutex_unlock(&lock);
 680 }
 681
 682 module_init(null_init);
 683 module_exit(null_exit);
 684
 685 MODULE_AUTHOR("Jens Axboe <jaxboe@fusionio.com>");
 686 MODULE_LICENSE("GPL");